Log anomaly detection with multidimensional features and decision-level fusion

We present LogMultiDFusion, a log anomaly detection framework that combines multidimensional feature modeling with decision-level fusion. In the feature modeling stage, we consider several candidate feature types and finally retain log semantics, event sequences, and event counts. For each retained channel, we train a detector that outputs normalized reconstruction errors and binary predictions. In the anomaly-detection stage, we compare two feature-level strategies— feature concatenation and cross-attention feature fusion—together with four decision-level fusion strategies (OR, AND, majority voting/k-of-n, and a learned logistic regression). Experiments on standard log datasets show that LogMultiDFusion consistently delivers the best overall F1-score and greater robustness across log types, confirming the effectiveness and practicality of combining multidimensional feature modeling with decision-level fusion.